Test Curation — Necessary vs Not

Companion to the generated Test Inventory and the Test Tiers. It records how the maintained test surface was curated through four lenses, in priority order:

1. Non-redundancy — cut where multiple layers cover the same thing.
2. Minimize runtime — push slow load/soak/stress to the highest tier.
3. Forward program — rust / qBittorrentBB / TrackMuleBB coverage, not just the frozen MFC app.
4. 0.7.3 correctness gate — what proves the shipped surface.

Verdict vocabulary: KEEP, WIRED-IN (was dormant, now gated), OPTIONAL (targeted-only / higher tier), TRIAGE (failing — fix before gating), CUT.

Headline findings

• The native layer's real problem was dormancy, not redundancy. 30 of 34 doctest suites (305 cases) ran by no tier or campaign — reachable only via test native --suite-name. Probing the built test binary showed most are fast (sub-100ms), seam-driven, and green.
• Three dormant suites were actually failing and nobody noticed because no tier ran them: fake_file_detector (1 failed of 129) and startup (1 of 61) were stale assertions — both fixed and wired — and divergence (8 of 431) is a deliberate 0.8.0 removal guard.
• The gate backbone parity was also red on a stale mule_list_ctrl search-column test (not updated after FEAT-118 merged Risk + Kad Confidence and FEAT-120 added the Extension column). Fixed to match MuleListCtrlViewPresets.h; the full 22-suite native gate is now green.
• The "duplicate profile" was a false positive. multi-client-p2p and multi-client-p2p-required share a suite list but differ in semantics — the -required variant enforces optional third-party clients as mandatory (live_e2e_suite.py branch on multi-client-p2p-required). Both are kept.
• Live-e2e tiering is already clean: the quick/fast set is entirely low-stress; every soak/stress/hammer/chaos suite is overnight-tier only (one orphan noted below).

Native layer — executed

19 dormant suites were verified green and fast against the built test binary and wired into test all (TEST_ALL_NATIVE_SUITES in test_runs.py), so they now gate at every tier:

async_dns_resolve, background_refresh, diagnostic_snapshot, fake_file_detector, kad-base, known_file_hash_open, packets, part_file_hash_launch, part_file_majority_name, process_launch, restart_app, search_trust_hint, server_connect, server_info, standby_prevention, startup, startup_storage, version_check_launch, windows_firewall_repair.

fake_file_detector and startup were failing on stale assertions; both were fixed, re-verified green against a fresh test build, and wired in with the rest.

Group	Verdict	Notes
parity (859 cases), web_api (87), protocol-parity (13)	KEEP	The 0.7.3 gate backbone.
community-core-divergence	KEEP	Orchestrated by community-core coverage (overnight).
The 19 suites above	WIRED-IN	Were dormant; verified green; now in test all.
divergence	KEEP DORMANT	A deliberate 0.8.0 scheduler-removal guard; red by design until the removal lands.
benchmark, pipeline, pipeline-benchmark	OPTIONAL	Performance, not correctness; stay targeted-only.
kad-broadband	KEEP DORMANT	fastkad_flow/kad_guards are <ClCompile> gated on Condition="Exists(...)" for KadPublishGuard.h/SafeKad.h (absent here) and an afximpl.h path; they are conditionally excluded from the build by design, so the suite registers zero cases. Not a gap.
Frozen MFC UI suites (*_keyboard_shortcuts, download_progress_bar, status_bar, pro_user_menu_copy, shared_dirs_tree_ctrl)	OPTIONAL	Frozen low-churn UI; targeted-only, out of the tiers.

Live-e2e layer

Group	Verdict	Notes
Fast set — preference-ui, shared-files-ui, config-stability-ui, shared-hash-ui, startup-diagnostics, shared-directories-rest, rest-api (+ auto-browse-live)	KEEP (quick/fast)	All stressClass = scenario.
Soak/stress/hammer/chaos + storage (10)	KEEP (overnight only)	Already overnight-tier profiles; none leak into quick/fast.
multi-client-p2p vs multi-client-p2p-required	KEEP (both)	Same suites, different evidence policy — intentional, not redundant.
shared-directory-browse-stress	WIRED-IN	Was an orphan (BUG-144 harness, self-contained fixture, unit-tested) in no profile. Added to stabilization-stress (overnight only); confirm on the next overnight run.
deterministic-two-client-transfer	KEEP (review)	Overlaps multi-client-p2p-matrix on local transfer; acceptable as a deterministic baseline.
Live-wire ARR (radarr/sonarr/prowlarr-emulebb)	KEEP (live-wire/release only)	Forward/controller surface; never quick/fast.

Python-harness layer

Group	Verdict	Notes
All modules	KEEP (quick/fast)	Fast unit tests, no app/network.
56 modules self-testing a live script	KEEP	Unit-vs-integration pairing, intentional. When a live suite/script is cut, cut its self-test too (mapping in the catalog selfTestsScript).

Triage of the failing suites

The failures were diagnosed against the built test binary (the relevant source had not changed since it was built, so they are real on current source — not stale-binary noise):

• divergence / scheduler_removal — deliberate, keep dormant. These are compile-time guards (#if __has_include("Scheduler.h"), #ifdef IDS_SCHEDULER, …) that assert the legacy scheduler is removed. It is still present, so they are red by design until the 0.8.0 legacy-surface removal lands. Correctly excluded from the 0.7.3 gate.
• startup / app_command_line — stale test, fixed and wired. The product was rebranded to emit "…canonical absolute eMuleBB base directory…" (AppCommandLineSeams.h), but the test still expected "eMule". Updated the expected string; a fresh test build confirmed 61/61 and the suite is now in test all.
• fake_file_detector — stale expectation, fixed and wired. BUG-150 added year-token stripping under broad release metadata, and the operator confirmed 2000 is the film's release year, so Sports Madness 2000 XviD 1080p correctly canonicalizes to sports madness (year dropped) while the title divergence is still flagged. Updated the assertion to expect sports madness; a fresh build confirmed 129/129 and the suite is now in test all.

Remaining / deferred

The native dormancy and the live orphan are now resolved. Two suites stay dormant by design (not gaps):

• divergence — red until the 0.8.0 scheduler removal lands.
• kad-broadband — conditionally excluded from the build until KadPublishGuard.h / SafeKad.h exist and the afximpl.h path gate is satisfied.

Performance suites (benchmark, pipeline, pipeline-benchmark) and the frozen MFC UI-shortcut suites remain targeted-only on purpose.

Stock/community parity is on-demand only. test protocol-parity (surface diff + goldens + live-diff vs the community baseline) and test community-core-coverage were removed from the overnight certification and from the emulebb-0.7.3 campaign. eMuleBB is protocol-stable and stock-compatible, so these baseline-build comparisons are run deliberately when protocol-adjacent code changes — not on every automated run. They remain fully available as the two standalone commands.

Regenerate the catalog after any change with python scripts/show-test-inventory.py --markdown and re-run python -m pytest tests/python/test_test_inventory.py.